xref: /netbsd-src/sys/kern/kern_condvar.c (revision 3007f1403a55612b606327e9e6f0e6e356b5bfa2)
1 /*	$NetBSD: kern_condvar.c,v 1.63 2023/11/02 10:31:55 martin Exp $	*/
2 
3 /*-
4  * Copyright (c) 2006, 2007, 2008, 2019, 2020, 2023
5  *     The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Andrew Doran.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30  * POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 /*
34  * Kernel condition variable implementation.
35  */
36 
37 #include <sys/cdefs.h>
38 __KERNEL_RCSID(0, "$NetBSD: kern_condvar.c,v 1.63 2023/11/02 10:31:55 martin Exp $");
39 
40 #include <sys/param.h>
41 
42 #include <sys/condvar.h>
43 #include <sys/cpu.h>
44 #include <sys/kernel.h>
45 #include <sys/lockdebug.h>
46 #include <sys/lwp.h>
47 #include <sys/sleepq.h>
48 #include <sys/syncobj.h>
49 #include <sys/systm.h>
50 
51 /*
52  * Accessors for the private contents of the kcondvar_t data type.
53  *
54  *	cv_opaque[0]	sleepq_t
55  *	cv_opaque[1]	description for ps(1)
56  *
57  * cv_opaque[0] is protected by the interlock passed to cv_wait() (enqueue
58  * only), and the sleep queue lock acquired with sleepq_hashlock() (enqueue
59  * and dequeue).
60  *
61  * cv_opaque[1] (the wmesg) is static and does not change throughout the life
62  * of the CV.
63  */
64 #define	CV_SLEEPQ(cv)		((sleepq_t *)(cv)->cv_opaque)
65 #define	CV_WMESG(cv)		((const char *)(cv)->cv_opaque[1])
66 #define	CV_SET_WMESG(cv, v) 	(cv)->cv_opaque[1] = __UNCONST(v)
67 
68 #define	CV_DEBUG_P(cv)	(CV_WMESG(cv) != nodebug)
69 #define	CV_RA		((uintptr_t)__builtin_return_address(0))
70 
71 static void		cv_unsleep(lwp_t *, bool);
72 static inline void	cv_wakeup_one(kcondvar_t *);
73 static inline void	cv_wakeup_all(kcondvar_t *);
74 
75 syncobj_t cv_syncobj = {
76 	.sobj_name	= "cv",
77 	.sobj_flag	= SOBJ_SLEEPQ_SORTED,
78 	.sobj_boostpri  = PRI_KERNEL,
79 	.sobj_unsleep	= cv_unsleep,
80 	.sobj_changepri	= sleepq_changepri,
81 	.sobj_lendpri	= sleepq_lendpri,
82 	.sobj_owner	= syncobj_noowner,
83 };
84 
85 static const char deadcv[] = "deadcv";
86 
87 /*
88  * cv_init:
89  *
90  *	Initialize a condition variable for use.
91  */
92 void
cv_init(kcondvar_t * cv,const char * wmesg)93 cv_init(kcondvar_t *cv, const char *wmesg)
94 {
95 
96 	KASSERT(wmesg != NULL);
97 	CV_SET_WMESG(cv, wmesg);
98 	sleepq_init(CV_SLEEPQ(cv));
99 }
100 
101 /*
102  * cv_destroy:
103  *
104  *	Tear down a condition variable.
105  */
106 void
cv_destroy(kcondvar_t * cv)107 cv_destroy(kcondvar_t *cv)
108 {
109 
110 	sleepq_destroy(CV_SLEEPQ(cv));
111 #ifdef DIAGNOSTIC
112 	KASSERT(cv_is_valid(cv));
113 	KASSERT(!cv_has_waiters(cv));
114 	CV_SET_WMESG(cv, deadcv);
115 #endif
116 }
117 
118 /*
119  * cv_enter:
120  *
121  *	Look up and lock the sleep queue corresponding to the given
122  *	condition variable, and increment the number of waiters.
123  */
124 static inline int
cv_enter(kcondvar_t * cv,kmutex_t * mtx,lwp_t * l,bool catch_p)125 cv_enter(kcondvar_t *cv, kmutex_t *mtx, lwp_t *l, bool catch_p)
126 {
127 	sleepq_t *sq;
128 	kmutex_t *mp;
129 	int nlocks;
130 
131 	KASSERT(cv_is_valid(cv));
132 	KASSERT(!cpu_intr_p());
133 	KASSERT((l->l_pflag & LP_INTR) == 0 || panicstr != NULL);
134 
135 	mp = sleepq_hashlock(cv);
136 	sq = CV_SLEEPQ(cv);
137 	nlocks = sleepq_enter(sq, l, mp);
138 	sleepq_enqueue(sq, cv, CV_WMESG(cv), &cv_syncobj, catch_p);
139 	mutex_exit(mtx);
140 	KASSERT(cv_has_waiters(cv));
141 	return nlocks;
142 }
143 
144 /*
145  * cv_unsleep:
146  *
147  *	Remove an LWP from the condition variable and sleep queue.  This
148  *	is called when the LWP has not been awoken normally but instead
149  *	interrupted: for example, when a signal is received.  Must be
150  *	called with the LWP locked.  Will unlock if "unlock" is true.
151  */
152 static void
cv_unsleep(lwp_t * l,bool unlock)153 cv_unsleep(lwp_t *l, bool unlock)
154 {
155 	kcondvar_t *cv __diagused;
156 
157 	cv = (kcondvar_t *)(uintptr_t)l->l_wchan;
158 
159 	KASSERT(l->l_wchan == (wchan_t)cv);
160 	KASSERT(l->l_sleepq == CV_SLEEPQ(cv));
161 	KASSERT(cv_is_valid(cv));
162 	KASSERT(cv_has_waiters(cv));
163 
164 	sleepq_unsleep(l, unlock);
165 }
166 
167 /*
168  * cv_wait:
169  *
170  *	Wait non-interruptably on a condition variable until awoken.
171  */
172 void
cv_wait(kcondvar_t * cv,kmutex_t * mtx)173 cv_wait(kcondvar_t *cv, kmutex_t *mtx)
174 {
175 	lwp_t *l = curlwp;
176 	int nlocks;
177 
178 	KASSERT(mutex_owned(mtx));
179 
180 	nlocks = cv_enter(cv, mtx, l, false);
181 	(void)sleepq_block(0, false, &cv_syncobj, nlocks);
182 	mutex_enter(mtx);
183 }
184 
185 /*
186  * cv_wait_sig:
187  *
188  *	Wait on a condition variable until a awoken or a signal is received.
189  *	Will also return early if the process is exiting.  Returns zero if
190  *	awoken normally, ERESTART if a signal was received and the system
191  *	call is restartable, or EINTR otherwise.
192  */
193 int
cv_wait_sig(kcondvar_t * cv,kmutex_t * mtx)194 cv_wait_sig(kcondvar_t *cv, kmutex_t *mtx)
195 {
196 	lwp_t *l = curlwp;
197 	int error, nlocks;
198 
199 	KASSERT(mutex_owned(mtx));
200 
201 	nlocks = cv_enter(cv, mtx, l, true);
202 	error = sleepq_block(0, true, &cv_syncobj, nlocks);
203 	mutex_enter(mtx);
204 	return error;
205 }
206 
207 /*
208  * cv_timedwait:
209  *
210  *	Wait on a condition variable until awoken or the specified timeout
211  *	expires.  Returns zero if awoken normally or EWOULDBLOCK if the
212  *	timeout expired.
213  *
214  *	timo is a timeout in ticks.  timo = 0 specifies an infinite timeout.
215  */
216 int
cv_timedwait(kcondvar_t * cv,kmutex_t * mtx,int timo)217 cv_timedwait(kcondvar_t *cv, kmutex_t *mtx, int timo)
218 {
219 	lwp_t *l = curlwp;
220 	int error, nlocks;
221 
222 	KASSERT(mutex_owned(mtx));
223 
224 	nlocks = cv_enter(cv, mtx, l, false);
225 	error = sleepq_block(timo, false, &cv_syncobj, nlocks);
226 	mutex_enter(mtx);
227 	return error;
228 }
229 
230 /*
231  * cv_timedwait_sig:
232  *
233  *	Wait on a condition variable until a timeout expires, awoken or a
234  *	signal is received.  Will also return early if the process is
235  *	exiting.  Returns zero if awoken normally, EWOULDBLOCK if the
236  *	timeout expires, ERESTART if a signal was received and the system
237  *	call is restartable, or EINTR otherwise.
238  *
239  *	timo is a timeout in ticks.  timo = 0 specifies an infinite timeout.
240  */
241 int
cv_timedwait_sig(kcondvar_t * cv,kmutex_t * mtx,int timo)242 cv_timedwait_sig(kcondvar_t *cv, kmutex_t *mtx, int timo)
243 {
244 	lwp_t *l = curlwp;
245 	int error, nlocks;
246 
247 	KASSERT(mutex_owned(mtx));
248 
249 	nlocks = cv_enter(cv, mtx, l, true);
250 	error = sleepq_block(timo, true, &cv_syncobj, nlocks);
251 	mutex_enter(mtx);
252 	return error;
253 }
254 
255 /*
256  * Given a number of seconds, sec, and 2^64ths of a second, frac, we
257  * want a number of ticks for a timeout:
258  *
259  *	timo = hz*(sec + frac/2^64)
260  *	     = hz*sec + hz*frac/2^64
261  *	     = hz*sec + hz*(frachi*2^32 + fraclo)/2^64
262  *	     = hz*sec + hz*frachi/2^32 + hz*fraclo/2^64,
263  *
264  * where frachi is the high 32 bits of frac and fraclo is the
265  * low 32 bits.
266  *
267  * We assume hz < INT_MAX/2 < UINT32_MAX, so
268  *
269  *	hz*fraclo/2^64 < fraclo*2^32/2^64 <= 1,
270  *
271  * since fraclo < 2^32.
272  *
273  * We clamp the result at INT_MAX/2 for a timeout in ticks, since we
274  * can't represent timeouts higher than INT_MAX in cv_timedwait, and
275  * spurious wakeup is OK.  Moreover, we don't want to wrap around,
276  * because we compute end - start in ticks in order to compute the
277  * remaining timeout, and that difference cannot wrap around, so we use
278  * a timeout less than INT_MAX.  Using INT_MAX/2 provides plenty of
279  * margin for paranoia and will exceed most waits in practice by far.
280  */
281 static unsigned
bintime2timo(const struct bintime * bt)282 bintime2timo(const struct bintime *bt)
283 {
284 
285 	KASSERT(hz < INT_MAX/2);
286 	CTASSERT(INT_MAX/2 < UINT32_MAX);
287 	if (bt->sec > ((INT_MAX/2)/hz))
288 		return INT_MAX/2;
289 	if ((hz*(bt->frac >> 32) >> 32) > (INT_MAX/2 - hz*bt->sec))
290 		return INT_MAX/2;
291 
292 	return hz*bt->sec + (hz*(bt->frac >> 32) >> 32);
293 }
294 
295 /*
296  * timo is in units of ticks.  We want units of seconds and 2^64ths of
297  * a second.  We know hz = 1 sec/tick, and 2^64 = 1 sec/(2^64th of a
298  * second), from which we can conclude 2^64 / hz = 1 (2^64th of a
299  * second)/tick.  So for the fractional part, we compute
300  *
301  *	frac = rem * 2^64 / hz
302  *	     = ((rem * 2^32) / hz) * 2^32
303  *
304  * Using truncating integer division instead of real division will
305  * leave us with only about 32 bits of precision, which means about
306  * 1/4-nanosecond resolution, which is good enough for our purposes.
307  */
308 static struct bintime
timo2bintime(unsigned timo)309 timo2bintime(unsigned timo)
310 {
311 
312 	return (struct bintime) {
313 		.sec = timo / hz,
314 		.frac = (((uint64_t)(timo % hz) << 32)/hz << 32),
315 	};
316 }
317 
318 /*
319  * cv_timedwaitbt:
320  *
321  *	Wait on a condition variable until awoken or the specified
322  *	timeout expires.  Returns zero if awoken normally or
323  *	EWOULDBLOCK if the timeout expires.
324  *
325  *	On entry, bt is a timeout in bintime.  cv_timedwaitbt subtracts
326  *	the time slept, so on exit, bt is the time remaining after
327  *	sleeping, possibly negative if the complete time has elapsed.
328  *	No infinite timeout; use cv_wait_sig instead.
329  *
330  *	epsilon is a requested maximum error in timeout (excluding
331  *	spurious wakeups).  Currently not used, will be used in the
332  *	future to choose between low- and high-resolution timers.
333  *	Actual wakeup time will be somewhere in [t, t + max(e, r) + s)
334  *	where r is the finest resolution of clock available and s is
335  *	scheduling delays for scheduler overhead and competing threads.
336  *	Time is measured by the interrupt source implementing the
337  *	timeout, not by another timecounter.
338  */
339 int
cv_timedwaitbt(kcondvar_t * cv,kmutex_t * mtx,struct bintime * bt,const struct bintime * epsilon __diagused)340 cv_timedwaitbt(kcondvar_t *cv, kmutex_t *mtx, struct bintime *bt,
341     const struct bintime *epsilon __diagused)
342 {
343 	struct bintime slept;
344 	unsigned start, end;
345 	int timo;
346 	int error;
347 
348 	KASSERTMSG(bt->sec >= 0, "negative timeout");
349 	KASSERTMSG(epsilon != NULL, "specify maximum requested delay");
350 
351 	/* If there's nothing left to wait, time out.  */
352 	if (bt->sec == 0 && bt->frac == 0)
353 		return EWOULDBLOCK;
354 
355 	/* Convert to ticks, but clamp to be >=1.  */
356 	timo = bintime2timo(bt);
357 	KASSERTMSG(timo >= 0, "negative ticks: %d", timo);
358 	if (timo == 0)
359 		timo = 1;
360 
361 	/*
362 	 * getticks() is technically int, but nothing special
363 	 * happens instead of overflow, so we assume two's-complement
364 	 * wraparound and just treat it as unsigned.
365 	 */
366 	start = getticks();
367 	error = cv_timedwait(cv, mtx, timo);
368 	end = getticks();
369 
370 	/*
371 	 * Set it to the time left, or zero, whichever is larger.  We
372 	 * do not fail with EWOULDBLOCK here because this may have been
373 	 * an explicit wakeup, so the caller needs to check before they
374 	 * give up or else cv_signal would be lost.
375 	 */
376 	slept = timo2bintime(end - start);
377 	if (bintimecmp(bt, &slept, <=)) {
378 		bt->sec = 0;
379 		bt->frac = 0;
380 	} else {
381 		/* bt := bt - slept */
382 		bintime_sub(bt, &slept);
383 	}
384 
385 	return error;
386 }
387 
388 /*
389  * cv_timedwaitbt_sig:
390  *
391  *	Wait on a condition variable until awoken, the specified
392  *	timeout expires, or interrupted by a signal.  Returns zero if
393  *	awoken normally, EWOULDBLOCK if the timeout expires, or
394  *	EINTR/ERESTART if interrupted by a signal.
395  *
396  *	On entry, bt is a timeout in bintime.  cv_timedwaitbt_sig
397  *	subtracts the time slept, so on exit, bt is the time remaining
398  *	after sleeping.  No infinite timeout; use cv_wait instead.
399  *
400  *	epsilon is a requested maximum error in timeout (excluding
401  *	spurious wakeups).  Currently not used, will be used in the
402  *	future to choose between low- and high-resolution timers.
403  */
404 int
cv_timedwaitbt_sig(kcondvar_t * cv,kmutex_t * mtx,struct bintime * bt,const struct bintime * epsilon __diagused)405 cv_timedwaitbt_sig(kcondvar_t *cv, kmutex_t *mtx, struct bintime *bt,
406     const struct bintime *epsilon __diagused)
407 {
408 	struct bintime slept;
409 	unsigned start, end;
410 	int timo;
411 	int error;
412 
413 	KASSERTMSG(bt->sec >= 0, "negative timeout");
414 	KASSERTMSG(epsilon != NULL, "specify maximum requested delay");
415 
416 	/* If there's nothing left to wait, time out.  */
417 	if (bt->sec == 0 && bt->frac == 0)
418 		return EWOULDBLOCK;
419 
420 	/* Convert to ticks, but clamp to be >=1.  */
421 	timo = bintime2timo(bt);
422 	KASSERTMSG(timo >= 0, "negative ticks: %d", timo);
423 	if (timo == 0)
424 		timo = 1;
425 
426 	/*
427 	 * getticks() is technically int, but nothing special
428 	 * happens instead of overflow, so we assume two's-complement
429 	 * wraparound and just treat it as unsigned.
430 	 */
431 	start = getticks();
432 	error = cv_timedwait_sig(cv, mtx, timo);
433 	end = getticks();
434 
435 	/*
436 	 * Set it to the time left, or zero, whichever is larger.  We
437 	 * do not fail with EWOULDBLOCK here because this may have been
438 	 * an explicit wakeup, so the caller needs to check before they
439 	 * give up or else cv_signal would be lost.
440 	 */
441 	slept = timo2bintime(end - start);
442 	if (bintimecmp(bt, &slept, <=)) {
443 		bt->sec = 0;
444 		bt->frac = 0;
445 	} else {
446 		/* bt := bt - slept */
447 		bintime_sub(bt, &slept);
448 	}
449 
450 	return error;
451 }
452 
453 /*
454  * cv_signal:
455  *
456  *	Wake the highest priority LWP waiting on a condition variable.  Must
457  *	be called with the interlocking mutex held or just after it has been
458  *	released (so the awoken LWP will see the changed condition).
459  */
460 void
cv_signal(kcondvar_t * cv)461 cv_signal(kcondvar_t *cv)
462 {
463 
464 	KASSERT(cv_is_valid(cv));
465 
466 	if (__predict_false(!LIST_EMPTY(CV_SLEEPQ(cv)))) {
467 		/*
468 		 * Compiler turns into a tail call usually, i.e. jmp,
469 		 * because the arguments are the same and no locals.
470 		 */
471 		cv_wakeup_one(cv);
472 	}
473 }
474 
475 /*
476  * cv_wakeup_one:
477  *
478  *	Slow path for cv_signal().  Deliberately marked __noinline to
479  *	prevent the compiler pulling it in to cv_signal(), which adds
480  *	extra prologue and epilogue code.
481  */
482 static __noinline void
cv_wakeup_one(kcondvar_t * cv)483 cv_wakeup_one(kcondvar_t *cv)
484 {
485 	sleepq_t *sq;
486 	kmutex_t *mp;
487 	lwp_t *l;
488 
489 	mp = sleepq_hashlock(cv);
490 	sq = CV_SLEEPQ(cv);
491 	if (__predict_true((l = LIST_FIRST(sq)) != NULL)) {
492 		KASSERT(l->l_sleepq == sq);
493 		KASSERT(l->l_mutex == mp);
494 		KASSERT(l->l_wchan == cv);
495 		sleepq_remove(sq, l, true);
496 	}
497 	mutex_spin_exit(mp);
498 }
499 
500 /*
501  * cv_broadcast:
502  *
503  *	Wake all LWPs waiting on a condition variable.  Must be called with
504  *	the interlocking mutex held or just after it has been released (so
505  *	the awoken LWP will see the changed condition).
506  */
507 void
cv_broadcast(kcondvar_t * cv)508 cv_broadcast(kcondvar_t *cv)
509 {
510 
511 	KASSERT(cv_is_valid(cv));
512 
513 	if (__predict_false(!LIST_EMPTY(CV_SLEEPQ(cv)))) {
514 		/*
515 		 * Compiler turns into a tail call usually, i.e. jmp,
516 		 * because the arguments are the same and no locals.
517 		 */
518 		cv_wakeup_all(cv);
519 	}
520 }
521 
522 /*
523  * cv_wakeup_all:
524  *
525  *	Slow path for cv_broadcast().  Deliberately marked __noinline to
526  *	prevent the compiler pulling it in to cv_broadcast(), which adds
527  *	extra prologue and epilogue code.
528  */
529 static __noinline void
cv_wakeup_all(kcondvar_t * cv)530 cv_wakeup_all(kcondvar_t *cv)
531 {
532 	sleepq_t *sq;
533 	kmutex_t *mp;
534 	lwp_t *l;
535 
536 	mp = sleepq_hashlock(cv);
537 	sq = CV_SLEEPQ(cv);
538 	while ((l = LIST_FIRST(sq)) != NULL) {
539 		KASSERT(l->l_sleepq == sq);
540 		KASSERT(l->l_mutex == mp);
541 		KASSERT(l->l_wchan == cv);
542 		sleepq_remove(sq, l, true);
543 	}
544 	mutex_spin_exit(mp);
545 }
546 
547 /*
548  * cv_has_waiters:
549  *
550  *	For diagnostic assertions: return non-zero if a condition
551  *	variable has waiters.
552  */
553 bool
cv_has_waiters(kcondvar_t * cv)554 cv_has_waiters(kcondvar_t *cv)
555 {
556 
557 	return !LIST_EMPTY(CV_SLEEPQ(cv));
558 }
559 
560 /*
561  * cv_is_valid:
562  *
563  *	For diagnostic assertions: return non-zero if a condition
564  *	variable appears to be valid.  No locks need be held.
565  */
566 bool
cv_is_valid(kcondvar_t * cv)567 cv_is_valid(kcondvar_t *cv)
568 {
569 
570 	return CV_WMESG(cv) != deadcv && CV_WMESG(cv) != NULL;
571 }
572